2.1. Summary of Previous Research
There were several operational applied modes in monitoring mine subsidence for the longwall excavation method of mining. For instance, in Jakarta (Indonesia), a combination of three methods was used together: leveling, GNSS, and interferometric synthetic aperture radar (InSAR). Observations have been running for a period of 28 years, yielding to the conclusion that temporal and spatial variations of measurements play the main role in understanding land subsidence behavior [
9].
Another comparison was done in Taiwan [
10], where an integrated model of ‘absolute’ and “relative” modes showed a significant impact on the dependability of subsidence. However, using two sets of GNSS solutions with the consistent differential mode of geoidal information between the benchmarks provided root mean square (RMS) difference of 5 cm.
A case study of deep mining subsidence [
11] provided an analysis of the combination of the probability-integral method with differential interferometric synthetic aperture radar (D-InSAR) technique. It was concluded that using the mining subsidence prediction model with D-InSAR could improve the accuracy of the quantitative estimation of deep mining subsidence rates and time threshold [
12].
In China, for the Nantun mining area, the subsidence down below the Zouji highway was introduced by a new methodology named probability integration model small baseline set (PIM-SBAS). This methodology is made with a combination of probability integration model (PIM), small baseline set (SBAS), and terraSAR-X (TSX) satellite images. Unfortunately, time-series interferometric synthetic aperture radar (TS-InSAR), where SBAS and TCPInSAR (temporarily coherent point InSAR) belong, did not reach expected results due to its capability to monitoring only linear subsidence where priority is given to deformation velocity. Comparison with classical leveling measurements showed that PIM-SBAS had better results than simple SBAS and TCPInSAR methods [
13].
Coal mining subsidence was possible to monitor as well by gravimetry instead of classical leveling surveys [
14]. The methodology was based on a negative linear correlation between gravity and height changes.
Airborne laser scanning (ALS) surveys were used for subsidence mapping in an Australian mine, where protentional benefits of this method were proven in the case study given in [
15].
Terrestrial laser scanner (TLS) has become an all-purpose tool for applications needing huge datasets. One of such applications is mine subsidence monitoring. The research on the utilization of TLS in mining applications is presented in [
16].
In Italy [
17] and in New South Wales, Australia [
18], subsidence was observed in three different ways, using unmanned aerial vehicles (UAV), total station, and laser imaging detection and ranging (LIDAR) measurements. Final results showed similar precision between the methods, but from experience, it was concluded that high-resolution images taken by UAV and processed by digital surface models (DSM), in comparison with two other techniques, were much more practical and faster.
For the Campine coal basin, Belgium, movement of the surface above the closed coal longwalls mine was studied for a period of 18 years using radar interferometry satellite data. The analysis was based on observation of the movement of downward and uplift of the coal basin. This effect was noticed in the majority of the closed coal mines and, in cases, where pumping water in the mine was stopped. The valuable conclusion said that even 40–60 years after closure, subsidence was still observable and occurring, of course, with a small rate. Using only satellite data for this type of analysis was absolutely adopted, but it might be supplemented by observing a larger area to better understand hydrological behavior, especially for flooding of underground zones [
19].
To estimate active subsidence over the coal mine Velenje, [
20] proposed a method where photogrammetric measurements obtained by an unmanned aerial vehicle (UAV) were introduced through a 3D point cloud. The observation area was divided into rectangular sectors, where also centroids of their planes were calculated. A model developed by the Faculty of Natural Sciences and Engineering (FNSE), in Ljubljana, was used to estimate the consolidation time, followed by cloud-to-cloud (C2C) analysis to identify the surface subsidence. This prognosis model gave only the prediction of subsidence from geometric and volume loss perspective.
2.2. Methods
As introduced previously through different examples, monitoring methodologies for any mining site have their own observation strategy based on environmental circumstances, availability of instrumental resources, accessibility to satellite data, the capability to manipulate by UAV, etc. Photogrammetry produces elevation models by UAV, which is a promising technology, due to its low costs regarding instrumentation and resources, as well as a huge amount of measured data. In the case study presented in this research, the overall accuracy of observing subsidence over Velenje coal mine using UAV was tested against reference methods, namely GNSS real-time kinematic (RTK) and total station (TS) tachymetry.
Our observation material consisted of measurements taken during 2017, where the test of congruence of UAV and traditional techniques were performed on five monitoring points (G1 to G5), located at the longwall site -80C over the Velenje coal mine. The heights of the monitoring points were determined by GNSS RTK and TS tachymetry. The monitoring points and the map of the terrain over the mine, as depicted in
Figure 5, were located over the central mining area. Such a decision was made since this is the area where the biggest displacements are expected: by computerized model [
21], where predicted final maximal vertical displacement on the surface above longwall -80C is approximately 10 m. All data were obtained from [
22]. Depending on the weather conditions, availability of access to the points, and the quality and acceptance of the measurements themselves, we identified a set of valid measurements collected during 2017.
The core of the research was to test the standard deviation of five control points’ heights obtained by UAV against those obtained by GNSS (and TS, where possible). UAV is the most promising method for massive data collection, especially in areas with limited accessibility. It is capable of creating high definition digital elevation models (DEMs), allowing various analyses of topography and artificial objects. Limiting issues for UAV photogrammetry, besides camera characteristics, are weather conditions (for example, strong winds or snow), errors in models, or wrong identification of surface points, just to mention some of the sources. Therefore, we chose certain five, well-established points, that could be easily identified in images. Those points’ coordinates were determined by GNSS and/or TS during each period when UAV campaigns were performed. The congruence of UAV-obtained heights with the ones determined by GNSS and TS was tested on height differences between the method under test (UAV photogrammetry) and the reference ones (GNSS RTK and TS tachymetry). Using that approach, the subsidence itself, as well as possible differences in coordinate system definitions, became negligible because only height differences were statistically tested on equality. It means that this research did not deal with subsidence predictions, but rather only with testing the values of the control points’ heights for particular measuring epochs. Therefore, there were no trends to incorporate in the statistical model. The choice of preferring GNSS RTK and TS tachymetry over geometric leveling (as the most accurate surveying method for determining heights) was a result of extremely hard environmental conditions, including wet and soft terrain with limited accessibility, where geometric leveling did not give acceptable results. There were few geometric leveling campaigns performed, but loop closures of leveling polygons, in all campaigns, did not fulfill quality assessment requirements. Therefore, this method was excluded from further investigation.
UAV measurements were performed using DJI Phantom 4 with DJI FC330 camera, in resolution 4000 px × 3000 px, focal length 3.61 mm, sensor size 6.2 mm × 4.6 mm, and pixel size 1.542 µm. The collected images were processed with 3D Survey aerial image processing software. The size of the obtained digital orthophoto (DOF) image was approximately 677 m × 743 m, with a pixel resolution of 4 cm.
Images are processed in 3DSurvey Mapping and Aerial Image Processing Software [
23]. The number of ground control points (GCPs) used for fitting DOF to the resulting coordinate system was 13–17, depending on the particular campaign. Fitting of DOF image to GCPs was around 5 cm for each measuring campaign, which was the value we adopted for the standard deviation of UAV-determined points’ heights, considering, also, flight plans, disposition, and density of ground control points. The bundle adjustment of the DOF models was performed following, so-called, global mode [
23], which allows for automatic checking of the produced model.
Figure 6 contains the DOFs for three selected UAV campaigns (1 February, 28 June 2017, and 15 January 2018). The images showed a decrease in access opportunities to the control points, which were located in the middle of the area. The third image in row (right), from
Figure 6, showed that the points G4 and G5 were covered by water, that is why data were not provided. Also, in some epochs, some of the points were under mud, which made the measurements unreliable or even impossible (for example, G3). Nevertheless, points’ disposition was carefully planned and selected, having in mind the safety of the field crew, as well as expected displacements.
GNSS observations were collected by LEICA Viva GS10 GNSS receiver, with a declared horizontal standard deviation of 8 mm + 1 ppm and a vertical standard deviation of 15 mm + 1 ppm, both in the RTK mode. The measuring sessions lasted 30 s per point. Having in mind measurement uncertainty of the method itself and transformation of the obtained coordinates to the state coordinate system, the standard deviation of GNSS RTK surveyed points’ heights was estimated to be 5 cm.
Also, three control height measurements were made by TS tachymetry. The measurements were performed by LEICA TCRA 1201+ (Heerbrugg, Switzerland) with a standard deviation of 1” for angles and 1 mm + 1.5 ppm for distances with a single reflective prism, which was the configuration utilized in this research. Although qualitative performance for this type of TS suggested measurements of sub-centimeter precision, it had to be realized that the area where the measurements were taken was wide and rich in topography. The mine control network consisted of a three-level network:
primary mine network, placed out of the influence zone,
measuring points network (for accessing detail), and
connecting network, which connected the networks 1 and 2
The final standard deviation of the points’ position in network 2 was estimated to be 5 cm. This value was completely sufficient for subsidence monitoring since the value of subsidence was 10 m per year. If we assumed that the maximum distance from the network point to a control point G-1 to G-5 was 500 m with the maximum vertical angle of ±10°, and estimated error of instrument’s height measurement of 1 mm, following the error propagation law, the resulting standard deviation of the observed point’s height would be:
with:
—standard deviation of a control point G-i (i = 1, 2, …, 5),
—standard deviation of a polygonometry point j,
—standard deviation of the height difference between P-j and C-i, and
—standard deviation of measured height of TS over the polygonometry point.
Since standard deviations of the measured height difference and height of the instrument were negligible to the standard deviation of the polygonometry point, the final value for the control point’s height surveyed by TS tachymetry was 5 cm, which was the value used further in this research.